Stimulus-specific delay activity in human primary visual cortex

Abstract

Working memory (WM) involves maintaining information in an on-line state. One emerging view is that information in WM is maintained via sensory recruitment, such that information is stored via sustained activity in the sensory areas that encode the to-be-remembered information. Using functional magnetic resonance imaging, we observed that key sensory regions such as primary visual cortex (V1) showed little evidence of sustained increases in mean activation during a WM delay period, though such amplitude increases have typically been used to determine whether a region is involved in on-line maintenance. However, a multivoxel pattern analysis of delay-period activity revealed a sustained pattern of activation in V1 that represented only the intentionally stored feature of a multifeature object. Moreover, the pattern of delay activity was qualitatively similar to that observed during the discrimination of sensory stimuli, suggesting that WM representations in V1 are reasonable "copies" of those evoked during pure sensory processing.

Fig. 1

The behavioral paradigm. On each trial, a sample stimulus that flickered at 5 Hz was presented for 1 s; observers were instructed to remember either the exact orientation or the exact color of this sample over the following 10-s delay period. After the delay, a test stimulus was presented, also for 1 s. The task was to indicate with a button press whether or not the test stimulus matched the sample stimulus on the indicated dimension. An exaggerated orientation-mismatch trial is depicted here for demonstration purposes. The test stimulus was followed by a 10-s intertrial interval (ITI).

Fig. 2

Mean amplitude of the blood-oxygenation-level-dependent (BOLD) response in primary visual cortex (V1) across the working memory (WM) delay period and intertrial interval (ITI). All time series were computed against a baseline of the activation level at Time 0. The vertical dotted lines highlight the onset of the sample stimulus at 0 s and the onset of the test stimulus at 11 s. The graphs at the top show results for (a) remember-orientation and (b) remember-color trials over a time window extending through 12 s post-stimulus, and the graphs at the bottom show results for (c) remember-orientation and (d) remember-color trials over a longer, 24-s temporal window. Note that because all the event-related time series were computed against a baseline of the respective activation level at Time 0 s, the second halves of the time series in (c) and (d) look slightly different from the ITI-evoked response in (a) and (b) even though they show the same data. Error bars represent ±1 SEM across observers.

Fig. 3

Feature-selective working memory (WM) modulations revealed by multivoxel pattern analysis. The graphs show classification accuracy as a function of the stimulus feature (color or orientation) being classified and whether the subject was instructed to remember orientation or color during the scan used as the basis for classification. Results are shown for (a) the WM delay period and (b) an analogous temporal interval following presentation of the test stimulus (i.e., during the intertrial interval, ITI). The horizontal lines at .5 accuracy highlight the level of chance performance. Error bars represent ±1 SEM across observers.

Fig. 4

Comparison of classification accuracy (collapsed across feature dimensions) in two time bins: 4 through 8 s poststimulus and 4 through 10 s poststimulus. The graph in (a) shows classification accuracy for remembered features (i.e., for orientation when subjects were remembering orientation and for color when subjects were remembering color) and for nonremembered features (e.g., for orientation when subjects were remembering color). The graph in (b) shows classification accuracy for the remembered (or relevant) feature based on data from the working memory (WM) delay period and based on data from the intertrial interval (ITI) following the test stimulus. (Note that following the test stimulus, the “remembered” feature did not need to be remembered any more and was simply the relevant feature for comparison with the sample.) Error bars represent ±1 SEM across observers.